Enhancing Fairness in Medical Image Classification: A Comparative Study of Convolutional Neural Networks and Adversarial Learning

Machine learning models have become increasingly prevalent in medical diagnosis and healthcare applications, yet these models often exhibit significant biases. This study investigates whether adversarial learning can effectively mitigate such biases and improve fairness in healthcare models. We compare a standard convolutional neural network (CNN) model with an adversarial learning approach on two datasets: the Diverse Dermatology Images Dataset (DDI), which includes demographic annotations, and the MNIST: HAM10000 dataset, which lacks such annotations. The adversarial learning framework incorporates a secondary model designed to challenge the primary model’s predictions based on sensitive attributes such as race and gender. By evaluating model performance and bias reduction across these datasets, we aim to determine the efficacy of adversarial learning in promoting more equitable outcomes. Our experiments reveal that while the CNNs exhibit varying levels of accuracy across datasets, adversarial learning significantly improves model performance and reduces racial bias. Specifically, the adversarial model outperforms the standard CNN in both datasets, achieving higher accuracy and demonstrating enhanced robustness against biases. This study underscores the potential of adversarial learning to foster more equitable machine learning models in healthcare for more optimized bias mitigation techniques and comprehensive fairness evaluation.